[Previous Article] [Next Article]
The two components of hurricane activity that we have studied are the seasonal numbers of hurricanes (H) and
seasonal numbers of intense hurricanes (I). To predict the number of hurricanes we use ordinary least squares regression
to estimate the number of tropical-only hurricanes, to which we add a constant seasonal average number of
baroclinically-influenced hurricanes. The regression used for prediction of tropical type hurricanes uses five predictors,
as described in Hess et al. (1995). These are the same as the basic, original predictors used by Gray et al. (1992) (see
the section in this issue containing Gray's forecasts). They include three QBO elements (50 and 30 mb zonal winds, and
their shear) and two African rainfall predictors (Gulf of Guinea and West Sahel). The direction of the relationships with
storm activity are: positive with the African rainfalls and the 50 mb zonal wind, negatively with the shear, and only
weakly with the 30 mb zonal wind when all terms are present. The baroclinically-influenced hurricanes are not
associated with these tropical-hurricane predictors; hence the constant adjustment to account for them in the total
prediction.
To predict the number of intense hurricanes (3 or more on the Saffir/Simpson scale) we use a nonlinear Poisson regression with maximum likelihood criterion. Details of this model are given in Elsner and Schmertmann (1993). The same five predictors used for tropical-only hurricanes are used for intense hurricanes. The November 1993 issue of this Bulletin briefly summarizes the reasoning behind the beneficial use of the Poisson for intense hurricane prediction. The Poisson equation looks like:
5
I = exp(o + i xi) for 5 predictors
i=1
Thus, in the Poisson the right hand side of a standard least squares regression equation is exponentiated.
In addition to basin-wide activity we are now predicting activity in four sub-basins of the Atlantic including the
Caribbean Sea, the Gulf of Mexico, the Southeast U.S. Coast (Cape Hatteras south to Key West) and the Northeast U.S.
coast (Cape Hatteras north to the Canadian border) (Lehmiller et al. 1997). We use logistic regression to predict
hurricane landfalls along the Northeast and Southeast coasts and the presence or absence of intense hurricanes in the
Gulf and Caribbean. As with the approach taken for basin-wide activity, we express the sub-basin forecasts in terms
of probabilities.
Logistic regression is a statistical model used to predict events in a yes/no framework by estimating coefficients
for several predictor variables. Here we use a maximum likelihood technique to obtain the coefficients. A logical
regression can be expressed as
exp (0 + 1x1 + 2x2 + ....... + pxp)
prob(yes) = -------------------------------------------------
1 + exp (0 + 1x1 + 2x2 + ....... + pxp)
where the 's are the coefficients on the p predictors x.
Unfortunately at present, for the early December forecasts the only model with significant skill above climatology is the model for Caribbean hurricanes. The Caribbean hurricane model uses only the two rainfall parameters (Gulf of Guinea and Sahel) as predictors.
Skills of the two prediction models for basin-wide storm activity were estimated using cross-validation over 45
years. Results indicate a mean absolute error of 1.5 hurricanes for the least squares regression model, compared to 1.7
for climatology forecasts. For the Poisson model forecasts for the number of intense hurricanes the mean absolute error
is 1.1, compared to 1.3 for climatology. Based on 1950-95 data, the prediction error for the logistic model for Caribbean
hurricanes, expressed as a cross-validated accuracy ratio, is 80.4% compared with a climatology ratio of 58.7%.
Using the least squares regression, 2.77 tropical-only hurricanes are forecast for the 1997 season. To this we add the updated seasonal average number of baroclinically-influenced hurricanes (2.9) and round the sum (5.67) to the nearest whole number to get a forecast of 6 Atlantic hurricanes for 1997. Using the Poisson model, the intense hurricane model forecast is presented below in the form of estimated probabilities for each possible number of intense hurricanes, for 1995 and 1996 (in retrospect) and for 1997.
Number of Intense Hurricanes
0 1 2 3 4 5+ mean
1995 Forecast Probability .047 .144 .220 .224 .171 .194 3.06
1996 Forecast Probability .259 .350 .236 .107 .036 .012 1.35
1997 Forecast Probability .110 .243 .268 .197 .109 .073 2.21
The 1997 probabilities are somewhat different from those of 1996, indicating an increased likelihood of intense
hurricane activity for 1997. More specifically, the Poisson model estimates a 71% chance of observing 1 to 3 intense
hurricanes in 1997. A forecast representing a rounded mean of the above distribution (mean = 2.21) calls for 2 intense
hurricanes during 1997 to develop. The logistic model for the occurrence of a hurricane in the Caribbean indicates
only a 31% chance of observing at least one in 1997. This is somewhat below the climatological probability.
The above forecasts will be updated prior to the start of the 1997 hurricane season.
The table below shows the regression coefficients for the equation for the number of tropical-only hurricanes (least
squares regression; first column), the number of intense hurricanes (Poisson model; second column), the number of
hurricanes in the Caribbean Basin (logistic regression model; third column), and the current predictor data used to form
all three of the 1997 forecasts (last column). The forecast models were estimated using 1950-96 data. The current
predictor data were obtained from Dr. Chris Landsea on December 2.
Regres Poisson Logistic
Coeff Coeff Coeff This
(Trop- (In- (Carib- Year's
Only tense bean Predictor
Predictor Term in Equation Hurr) Hurr) Hurr) Value
---------------------------------------------------------------------------------------------------
Constant 4.485 1.250 1.123 ---
Aug-Nov '96 Gulf of Guinea rainfall 2.762 0.614 3.440 -0.54 sd
Aug-Sep '96 West Sahel rainfall 0.780 0.304 1.201 -0.05 sd
50mb zonal wind, 10ON, fcst Sep '97 0.025 0.034 --- -3 ms-1
30mb zonal wind, 10ON, fcst Sep '97 -0.012 -0.015 --- -2 ms-1
Absolute shear from two above winds -0.133 -0.039 --- 1 ms-1
---------------------------------------------------------------------------------------------------
Verification of the 1996 Hurricane Season Forecasts
Using the above statistical models, the following forecasts were made, compared with the observations:
Early Early Observed
Dec. Aug. (% of norm)
----- ----- -----------
Hurricanes 6 7 9 (155%)
Intense Hurric. 1 2 6 (261%)
The total activity was above normal and intense hurricane activity was much above average, which busted our
forecasts. There were 6 intense hurricanes (although three of them reached only the minimum threshold of 50 m/s); the
early December Poisson model estimated only a 1.2% chance of 5 or more intense hurricanes.
We speculate that the bust was partly due to the breakdown in the normal persistence of the Sahelian rainfall
anomaly: Aug-Sep 1995 was quite dry, suggesting a dry Aug-Sep 1996 and thus an inactive 1996 storm season. But
Sahelian rainfall in Aug-Sep 1996 turned out to be near normal.
The 1996 continued a trend from 1995 of a majority of the hurricanes originating from African tropical waves.
Six of the 9 in 1996 visited the Caribbean Sea--the most since 1916. Using the objective classification rules of Elsner
et al. (1996) for identifying hurricanes of purely tropical origin versus those with mid-latitude baroclinic influences, 7
of the 9 were tropical-only. However, careful subjective analysis indicates that of the two that were classified as having
baroclinic influence (Lili and Marco), the classification of Lili appears uncertain and that of Marco is incorrect (i.e. it
should have been called a tropical-only storm).
Although we did not issue a hurricane location forecast last December, our location forecasts from early August
verified fairly well for the balance of the 1996 season. Of the 4 location forecasts issued, 3 proved to be successful. In
particular, we accurately predicted that there would be no intense hurricane activity in the Gulf of Mexico or the
Caribbean Sea (model probability predictions were 13% and 40%, respectively). The Caribbean forecast was particularly
successful in that out of the 4 hurricanes that occurred in the Caribbean after August 1, none reached intense strength.
We were unsuccessful in forecasting U.S. east coast landfalls. Although we accurately predicted that no hurricanes
would make landfall on the Northeast Coast (model prediction was 1%), the model for the Southeast Coast gave a 9%
probability of landfall. Thus, we forecast that no hurricanes would hit the Southeast Coast after early August.
Unfortunately, hurricane Fran invalidated this forecast.
The model for each location has an expected accuracy of approximately 80% (Lehmiller et al. 1997). In 1996, 3
of the 4 regional models provided accurate predictions. Overall, then, the location models performed about as expected,
with long-run expectations slightly higher than the 1996 performance.
Acknowledgments: We are grateful to Mr. J. Weigel for his help with the 1996 hurricane statistics. Partial support
for this work came from the Risk Prediction Initiative (RPI) of the Bermuda Biological Station for Research (BBSR).
Elsner, J.B. and C.P. Schmertmann, 1993: Improving extended-range seasonal predictions of intense Atlantic
hurricane activity. Wea. Forecasting, 8, 345-351.
Elsner, J.B., G.S. Lehmiller and T.B. Kimberlain, 1996: Objective classification of Atlantic basin hurricanes. J.
Climate, 9, 2880-2889.
Gray, W.M., C.W. Landsea, P.W. Mielke, and K.J. Berry, 1992: Predicting Atlantic seasonal hurricane activity 6-11 months in advance. Wea. Forecasting, 7, 440-455.
Hess, J.C., J.B. Elsner and N.E. LaSeur, 1995: Improving seasonal hurricane predictions for the Atlantic Basin. Wea. Forecasting, 10, 425-432.
Lehmiller, G.S., T.B. Kimberlain and J.B. Elsner, 1997: Seasonal prediction models for North Atlantic basin
hurricane location. Mon. Wea. Rev., 125, accepted.